Compute & Infrastructure

Beijing has announced the Space Computing Industry Innovation Center, a state-directed consortium bringing together Chinese rocket manufacturers, satellite operators, chip designers, and AI laboratories to develop grid-independent, space-based AI data centre infrastructure. The announcement, reported by Tom's Hardware, was timed a week before SpaceX's AI1 reveal, and the proximity is unlikely to be coincidental. The framing is explicitly competitive with Musk's commercial space compute ambitions.

The strategic logic is layered. Space-based compute sidesteps terrestrial power grid constraints that are slowing data centre buildout globally, avoids Western export controls on advanced semiconductors by developing indigenous chip capacity within a closed alliance, and positions China to claim sovereign compute infrastructure that is physically beyond jurisdictional reach. Whether the technical and cost challenges of operating high-density compute in orbit are solvable at scale remains unproven — this is firmly in the announced-plan category, not confirmed capacity. But the institutional architecture being assembled is real and signals Beijing's willingness to pursue asymmetric infrastructure strategies.

A preprint study, covered by Tom's Hardware, alleges that Pearl's decentralised AI-mining network is consuming the equivalent of 320,000 RTX 3090-class GPUs and 112 MW of power while performing random matrix operations with no verified AI output. The practical consequence is a 38% jump in GPU rental costs across the market — a direct supply signal that affects legitimate AI training and inference operators who rely on spot and rental GPU capacity.

This matters beyond the fraud angle. If a meaningful fraction of GPU rental supply is being absorbed by workloads that produce no useful computation, demand-side metrics used by analysts and hyperscalers to project capacity requirements are contaminated. Infrastructure planners relying on GPU utilisation data to size future procurement could be systematically overestimating real AI workload density, or alternatively underestimating how much additional capacity is needed to serve genuine demand. The study is a preprint and Pearl disputes the characterisation, so the precise figures should be treated as contested — but the structural dynamic of financialised GPU demand inflating spot markets is corroborated by the rental price movement.

A sponsored analysis from Data Center Dynamics argues that Battery Energy Storage Systems are transitioning from a resilience tool to a structural component of AI data centre power architecture. The core driver is speed-to-power: grid connection timelines in key markets now run to multi-year queues, and operators deploying BESS can draw down stored capacity to begin operations while permanent grid connections are finalised. For hyperscalers racing to monetise capacity, this compresses the revenue gap between construction completion and live workloads.

The secondary function is load shaping — BESS allows data centres to charge during off-peak periods and discharge during peak AI training runs, which both reduces grid stress and provides a degree of insulation from demand-surge pricing. The caveat is that BESS at data centre scale requires substantial lithium supply chain exposure, and the economics depend heavily on local electricity tariff structures and grid operator willingness to credit stored energy. This is a confirmed and accelerating trend, not speculative — major data centre operators including hyperscalers have announced BESS deployments in the US, UK, and Europe over the past 18 months.

Intel and AMD have jointly published the ACE instruction set extensions for x86, designed to make matrix multiplication — the core operation in neural network inference — more power- and density-efficient on general-purpose CPUs, as reported by Tom's Hardware. Separately, Microsoft is testing Copilot+ AI features on discrete GPUs rather than requiring dedicated NPUs, expanding the addressable device base for local AI workloads.

Taken together, these moves reflect a structural shift in where AI inference runs: away from centralised GPU clusters and toward the existing installed base of client hardware. For data centre operators and cloud providers, this is a long-run demand moderation signal for inference capacity — if a meaningful fraction of AI assistant workloads can execute locally on CPUs or consumer GPUs, the inference traffic that would otherwise hit cloud endpoints is reduced. The timeline for this to affect hyperscaler capacity planning is long, as software ecosystems and device refresh cycles operate on multi-year horizons, but the directional pressure is established.

With under a year remaining in office, Emmanuel Macron is positioning France's G7 presidency around AI infrastructure commitments and data centre investment pledges, according to Bloomberg. The analysis notes that the funding is characterised as 'fickle' — a signal that a significant portion of the headline investment figures are conditional, announced, or dependent on private sector follow-through that has not yet materialised as confirmed construction or capacity.

This fits a wider pattern in European sovereign compute strategy: large headline numbers attached to AI infrastructure summits that, on closer inspection, blend committed public funds, conditional private pledges, and aspirational targets. The US government's parallel push, flagged in Bloomberg reporting on FedHIVE CEO Michael Cardaci's comments, is moving with more urgency on data sovereignty and compute compliance — keeping US technology within US borders as a national security posture — which creates additional friction for European cloud infrastructure that relies on US hyperscaler capacity.